Merge pull request #2261 from Molecule-AI/fix/harness-cleanup-failed-event

harness: SaaS routing + provider-agnostic config for RFC #2251 measurement
This commit is contained in:
hongmingwang-moleculeai 2026-04-29 05:35:43 +00:00 committed by GitHub
commit a18d116606
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 427 additions and 0 deletions

47
scripts/README.md Normal file
View File

@ -0,0 +1,47 @@
# scripts/
Operational and one-off scripts for molecule-core. Most are
self-documenting — see the header comments in each file.
## RFC #2251 coordinator task-bound harnesses
There are three related scripts; pick the right one:
| Script | Purpose | Targets |
|---|---|---|
| `measure-coordinator-task-bounds.sh` | **Canonical** v1 harness for the RFC #2251 / Issue 4 reproduction. Provisions a PM coordinator + Researcher child via `claude-code-default` + `langgraph` templates, sends a synthesis-heavy A2A kickoff, observes elapsed time + heartbeat trace. | OSS-shape platform — localhost or any `/workspaces`-shaped endpoint. Has tenant/admin-token guards for non-localhost runs. |
| `measure-coordinator-task-bounds-runner.sh` | Generalised runner for the same measurement contract but with **arbitrary template + secret + model combinations** (Hermes/MiniMax, etc.). Useful for cross-runtime variants without modifying the canonical harness. | Same as above (local or SaaS via `MODE=saas`). |
| `measure-coordinator-task-bounds.sh` (in [molecule-controlplane](https://github.com/Molecule-AI/molecule-controlplane)) | **Production-shape** variant that bootstraps a real staging tenant via `POST /cp/admin/orgs`, then runs the same measurement against `<slug>.staging.moleculesai.app`. | Staging controlplane only — refuses to run against production. |
See `reference_harness_pair_pattern` (auto-memory) for when to use which
and the cross-repo design rationale.
### Common safety pattern across all three
- **Cleanup trap** on EXIT/INT/TERM auto-deletes provisioned resources.
- **`DRY_RUN=1`** prints plan + auth fingerprint, exits before any
state mutation. Run this before pointing at staging or any shared
infrastructure.
- **Non-target guard** refuses arbitrary endpoints (the controlplane
variant is locked to `staging-api.moleculesai.app`; the OSS variant
requires explicit auth + tenant scoping for non-localhost PLATFORM).
- **Cleanup failures emit `cleanup_*_failed` events** with remediation
hints; no silenced curl. ADMIN_TOKEN expiring mid-run surfaces as a
structured event rather than a silent leak.
### Heartbeat trace caveat
If `heartbeat_trace.raw == "<endpoint_unavailable>"`, the per-workspace
`/heartbeat-history` endpoint isn't wired on the target build — the
bound measurement is INCONCLUSIVE on the platform-ceiling question.
Either wire the endpoint or replace with the equivalent Datadog query.
## Other scripts
- `cleanup-rogue-workspaces.sh` — emergency teardown for leaked
workspaces. Prompts for confirmation. Pair with the harnesses if a
cleanup trap fails (see `cleanup_*_failed` events).
- `canary-smoke.sh` — quick smoke test for canary releases.
- `dev-start.sh` — local-dev platform bring-up.
The rest are self-documenting in their header comments.

View File

@ -0,0 +1,273 @@
#!/usr/bin/env bash
# Standalone runner for Issue 4 reproduction (RFC #2251) — exists alongside
# `measure-coordinator-task-bounds.sh` to support arbitrary template + secret
# combinations without modifying the canonical harness. The canonical harness
# stays focused on its v1 contract (claude-code-default + langgraph + OpenRouter);
# this runner wraps the same workspace-server API calls but takes everything as
# env-var inputs so a Hermes/MiniMax run can share the measurement code path.
#
# Two routing modes:
# MODE=local (default) — direct workspace-server API
# MODE=saas — placeholder; populates same vars but expects
# PLATFORM=<tenant-subdomain> with X-Tenant-Id +
# Authorization headers from CP_ADMIN_API_TOKEN
#
# Required env:
# PLATFORM workspace-server base URL (default http://localhost:8080)
# PM_TEMPLATE template slug for coordinator
# CHILD_TEMPLATE template slug for researcher child
# SECRET_NAME workspace_secrets key (e.g. MINIMAX_API_KEY)
# SECRET_VALUE the secret value (or read from $SECRET_NAME if unset)
#
# Optional:
# MODEL PUT /workspaces/:id/model after provision
# SYNTHESIS_DEPTH=3 number of delegation rounds in the kickoff task
# A2A_TIMEOUT=600 ceiling on measurement-side wait (seconds)
# KEEP_WORKSPACES=0 skip cleanup-on-exit when 1 (for log inspection)
# MODE=local|saas local-dev vs SaaS routing posture
# CP_ADMIN_API_TOKEN required when MODE=saas; sent as Authorization bearer
# TENANT_ID required when MODE=saas; sent as X-Tenant-Id
#
# Output: NDJSON event stream on stdout + a human summary on stderr.
#
set -euo pipefail
PLATFORM="${PLATFORM:-http://localhost:8080}"
MODE="${MODE:-local}"
PM_TEMPLATE="${PM_TEMPLATE:?PM_TEMPLATE is required (e.g. claude-code-default, hermes)}"
CHILD_TEMPLATE="${CHILD_TEMPLATE:?CHILD_TEMPLATE is required}"
SECRET_NAME="${SECRET_NAME:?SECRET_NAME is required (e.g. MINIMAX_API_KEY)}"
MODEL="${MODEL:-}"
SYNTHESIS_DEPTH="${SYNTHESIS_DEPTH:-3}"
A2A_TIMEOUT="${A2A_TIMEOUT:-600}"
KEEP_WORKSPACES="${KEEP_WORKSPACES:-0}"
# SaaS-mode auth chain: workspace-server (per-tenant Go binary on EC2)
# requires BOTH headers:
# Authorization: Bearer <tenant-admin-token> (per-tenant secret)
# X-Molecule-Org-Id: <org-uuid> (TenantGuard middleware)
# The tenant-admin-token is provisioned by controlplane and retrievable via:
# GET /cp/admin/orgs/<slug>/admin-token (CP_ADMIN_API_TOKEN bearer-gated)
# The runner can either:
# 1. Take ORG_SLUG + CP_ADMIN_API_TOKEN and fetch the tenant token itself, or
# 2. Take ORG_ID + TENANT_ADMIN_TOKEN directly.
ORG_ID="${ORG_ID:-}"
ORG_SLUG="${ORG_SLUG:-}"
TENANT_ADMIN_TOKEN="${TENANT_ADMIN_TOKEN:-}"
CP_ADMIN_API_TOKEN="${CP_ADMIN_API_TOKEN:-}"
CP_API_URL="${CP_API_URL:-https://staging-api.moleculesai.app}"
# Resolve secret value: ${SECRET_VALUE} > $${SECRET_NAME} > error.
SECRET_VALUE="${SECRET_VALUE:-}"
if [ -z "$SECRET_VALUE" ]; then
SECRET_VALUE="$(printenv "$SECRET_NAME" 2>/dev/null || true)"
fi
[ -n "$SECRET_VALUE" ] || { echo "ERROR: set \$$SECRET_NAME or \$SECRET_VALUE" >&2; exit 1; }
# SaaS-mode preflight + format validation.
# Validating ORG_ID + ORG_SLUG client-side gives an actionable error
# before the request hits TenantGuard's intentionally-opaque 404
# (which doesn't tell the operator whether the slug is wrong, the
# UUID is wrong, or auth is wrong).
if [ "$MODE" = "saas" ]; then
[ -n "$ORG_ID" ] || { echo "ERROR: MODE=saas requires ORG_ID (the org UUID)" >&2; exit 1; }
case "$ORG_ID" in
[0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f]-[0-9a-f][0-9a-f][0-9a-f][0-9a-f]-[0-9a-f][0-9a-f][0-9a-f][0-9a-f]-[0-9a-f][0-9a-f][0-9a-f][0-9a-f]-[0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f][0-9a-f]) ;;
*) echo "ERROR: ORG_ID must be a UUID (got '$ORG_ID')" >&2; exit 1;;
esac
if [ -n "$ORG_SLUG" ]; then
case "$ORG_SLUG" in
*[!a-z0-9-]* | -* | *-) echo "ERROR: ORG_SLUG must match ^[a-z0-9][a-z0-9-]*[a-z0-9]\$ (got '$ORG_SLUG')" >&2; exit 1;;
esac
fi
if [ -z "$TENANT_ADMIN_TOKEN" ]; then
[ -n "$ORG_SLUG" ] || { echo "ERROR: MODE=saas needs TENANT_ADMIN_TOKEN or ORG_SLUG (to fetch it via CP)" >&2; exit 1; }
[ -n "$CP_ADMIN_API_TOKEN" ] || { echo "ERROR: ORG_SLUG path needs CP_ADMIN_API_TOKEN to fetch tenant token from $CP_API_URL" >&2; exit 1; }
TENANT_ADMIN_TOKEN=$(curl -s -H "Authorization: Bearer $CP_ADMIN_API_TOKEN" \
"$CP_API_URL/cp/admin/orgs/$ORG_SLUG/admin-token" \
| python3 -c "import sys,json; print(json.load(sys.stdin).get('admin_token',''))" 2>/dev/null || echo "")
[ -n "$TENANT_ADMIN_TOKEN" ] || { echo "ERROR: failed to resolve tenant admin token via $CP_API_URL/cp/admin/orgs/$ORG_SLUG/admin-token" >&2; exit 1; }
fi
fi
ts() { date -u +%Y-%m-%dT%H:%M:%S.%3NZ 2>/dev/null || date -u +%Y-%m-%dT%H:%M:%SZ; }
emit() { printf '{"ts":"%s","event":"%s","data":%s}\n' "$(ts)" "$1" "${2:-null}"; }
api() {
local args=()
if [ "$MODE" = "saas" ]; then
args+=(-H "Authorization: Bearer $TENANT_ADMIN_TOKEN")
args+=(-H "X-Molecule-Org-Id: $ORG_ID")
fi
curl -s ${args[@]+"${args[@]}"} "$@"
}
PM_ID=""
CHILD_ID=""
cleanup() {
local rc=$?
set +e
if [ "$KEEP_WORKSPACES" = "1" ]; then
emit "cleanup_skipped" "{\"reason\":\"KEEP_WORKSPACES=1\",\"pm_id\":\"$PM_ID\",\"child_id\":\"$CHILD_ID\"}"
return $rc
fi
for id in "$CHILD_ID" "$PM_ID"; do
[ -z "$id" ] && continue
code=$(api -o /dev/null -w '%{http_code}' -X DELETE "$PLATFORM/workspaces/$id" 2>/dev/null || echo "curl_err")
if [ "$code" = "200" ] || [ "$code" = "204" ] || [ "$code" = "404" ]; then
emit "cleanup_deleted" "{\"workspace_id\":\"$id\",\"http_code\":\"$code\"}"
else
emit "cleanup_failed" "{\"workspace_id\":\"$id\",\"http_code\":\"$code\"}"
fi
done
return $rc
}
trap cleanup EXIT INT TERM
emit "run_started" "{\"platform\":\"$PLATFORM\",\"mode\":\"$MODE\",\"pm_template\":\"$PM_TEMPLATE\",\"child_template\":\"$CHILD_TEMPLATE\",\"model\":\"$MODEL\",\"secret_name\":\"$SECRET_NAME\",\"synthesis_depth\":$SYNTHESIS_DEPTH,\"a2a_timeout_secs\":$A2A_TIMEOUT}"
# ---- Provision via JSON-encoded bodies (defends against templates/values
# with embedded shell-special chars). ----
pm_body=$(python3 -c '
import json, sys
print(json.dumps({"name":"PM","role":"Coordinator — delegates and synthesizes","tier":2,"template":sys.argv[1]}))' "$PM_TEMPLATE")
child_body=$(python3 -c '
import json, sys
print(json.dumps({"name":"Researcher","role":"Returns short research findings","tier":2,"template":sys.argv[1]}))' "$CHILD_TEMPLATE")
secret_body=$(python3 -c '
import json, sys
print(json.dumps({"key":sys.argv[1],"value":sys.argv[2]}))' "$SECRET_NAME" "$SECRET_VALUE")
emit "provisioning_pm" "{\"template\":\"$PM_TEMPLATE\"}"
R=$(api -X POST "$PLATFORM/workspaces" -H 'Content-Type: application/json' -d "$pm_body")
PM_ID=$(echo "$R" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null || echo "")
[ -n "$PM_ID" ] || { echo "ERROR: PM create failed — response: $R" >&2; exit 1; }
emit "pm_provisioned" "{\"workspace_id\":\"$PM_ID\"}"
emit "provisioning_child" "{\"template\":\"$CHILD_TEMPLATE\"}"
R=$(api -X POST "$PLATFORM/workspaces" -H 'Content-Type: application/json' -d "$child_body")
CHILD_ID=$(echo "$R" | python3 -c "import sys,json; print(json.load(sys.stdin).get('id',''))" 2>/dev/null || echo "")
[ -n "$CHILD_ID" ] || { echo "ERROR: child create failed — response: $R" >&2; exit 1; }
emit "child_provisioned" "{\"workspace_id\":\"$CHILD_ID\"}"
api -X PATCH "$PLATFORM/workspaces/$CHILD_ID" -H 'Content-Type: application/json' \
-d "{\"parent_id\":\"$PM_ID\"}" > /dev/null
# Seed secret on BOTH workspaces. Hermes/MiniMax both sides need it; templates
# that ignore unknown env vars treat extras as no-op.
for id in "$PM_ID" "$CHILD_ID"; do
api -X POST "$PLATFORM/workspaces/$id/secrets" -H 'Content-Type: application/json' -d "$secret_body" > /dev/null
done
emit "secrets_seeded" "{\"key\":\"$SECRET_NAME\",\"workspaces\":[\"$PM_ID\",\"$CHILD_ID\"]}"
if [ -n "$MODEL" ]; then
model_body=$(python3 -c 'import json,sys; print(json.dumps({"model":sys.argv[1]}))' "$MODEL")
for id in "$PM_ID" "$CHILD_ID"; do
api -X PUT "$PLATFORM/workspaces/$id/model" -H 'Content-Type: application/json' -d "$model_body" > /dev/null
done
emit "model_set" "{\"model\":\"$MODEL\",\"workspaces\":[\"$PM_ID\",\"$CHILD_ID\"]}"
fi
# ---- Wait for both online ----
WAIT_ONLINE_SECS="${WAIT_ONLINE_SECS:-180}"
wait_online() {
local id="$1" label="$2"
# Round up so a non-multiple-of-3 budget waits at least the requested
# seconds (200 → 67 polls × 3s = 201s, not 198s).
local polls=$(( (WAIT_ONLINE_SECS + 2) / 3 ))
local last_status=""
for i in $(seq 1 "$polls"); do
s=$(api "$PLATFORM/workspaces/$id" | python3 -c "import sys,json; print(json.load(sys.stdin).get('status',''))" 2>/dev/null || echo "")
if [ "$s" != "$last_status" ]; then
emit "status_change" "{\"workspace\":\"$label\",\"from\":\"$last_status\",\"to\":\"$s\",\"poll\":$i}"
last_status="$s"
fi
[ "$s" = "online" ] && { emit "online" "{\"workspace\":\"$label\",\"after_polls\":$i,\"after_secs\":$((i * 3))}"; return 0; }
[ "$s" = "failed" ] && { emit "failed" "{\"workspace\":\"$label\"}"; return 1; }
sleep 3
done
emit "online_timeout" "{\"workspace\":\"$label\",\"last_status\":\"$last_status\",\"waited_secs\":$WAIT_ONLINE_SECS}"
return 1
}
wait_online "$PM_ID" "PM" || exit 2
wait_online "$CHILD_ID" "child" || exit 2
# ---- Build a synthesis-heavy kickoff task ----
TASK="You are coordinating a research analysis. Delegate $SYNTHESIS_DEPTH separate sub-questions to the Researcher (one at a time, sequentially — wait for each response before sending the next), then synthesize all findings into a single coherent report. Sub-questions: (a) historical context of distributed consensus, (b) modern Byzantine-fault-tolerant protocols, (c) practical trade-offs between Raft and Paxos. After all delegations complete, write a 600-word synthesis comparing the three responses and drawing one cross-cutting insight. Do not respond until the synthesis is complete."
# ---- A2A kickoff round-trip ----
emit "a2a_kickoff_sent" "{\"to\":\"$PM_ID\",\"task_chars\":${#TASK}}"
START_NS=$(python3 -c 'import time; print(int(time.time_ns()))')
a2a_body=$(python3 -c '
import json, sys
print(json.dumps({"method":"message/send","params":{"message":{"role":"user","parts":[{"type":"text","text":sys.argv[1]}]}}}))' "$TASK")
RESP=$(api --max-time "$A2A_TIMEOUT" -X POST "$PLATFORM/workspaces/$PM_ID/a2a" \
-H "Content-Type: application/json" -d "$a2a_body" || echo "<curl_failed_or_timed_out>")
END_NS=$(python3 -c 'import time; print(int(time.time_ns()))')
ELAPSED_SECS=$(python3 -c "print(round(($END_NS - $START_NS) / 1e9, 2))")
emit "a2a_response_observed" "{\"elapsed_secs\":$ELAPSED_SECS,\"response_chars\":${#RESP},\"response_head\":$(python3 -c "import json,sys; print(json.dumps(sys.argv[1][:200]))" "$RESP")}"
# ---- Heartbeat trace ----
# `/workspaces/:id/heartbeat-history` is a local-dev workspace-server
# route — on tenant deployments the platform's :8080 fallback proxies
# any unmatched path to the canvas Next.js, so this 404s with 28KB of
# HTML rather than a clean error. Skip the fetch entirely in SaaS mode
# and emit an explicit placeholder instead of polluting the event log
# with HTML.
emit "fetching_heartbeat_trace" "{\"mode\":\"$MODE\"}"
if [ "$MODE" = "saas" ]; then
emit "heartbeat_trace" "{\"raw\":\"<skipped: heartbeat-history endpoint unavailable in SaaS — see scripts/README.md\>\"}"
else
HB=$(api "$PLATFORM/workspaces/$PM_ID/heartbeat-history?since_secs=$A2A_TIMEOUT" 2>&1 || echo "<endpoint_unavailable>")
emit "heartbeat_trace" "{\"raw\":$(python3 -c "import json,sys; print(json.dumps(sys.argv[1]))" "$HB")}"
fi
# ---- rfc2251_phase log lines from the workspace container ----
# Local Docker provisioner: workspace container name is workspace-<id>.
# SaaS: container is on EC2 — skip log capture, fall back to heartbeat only.
if [ "$MODE" = "local" ] && command -v docker >/dev/null 2>&1; then
for id in "$PM_ID"; do
container=$(docker ps --filter "name=workspace-$id" --format '{{.Names}}' | head -1)
if [ -n "$container" ]; then
phase_log=$(docker logs --since "${A2A_TIMEOUT}s" "$container" 2>&1 | grep 'rfc2251_phase=' || echo "<no rfc2251_phase log lines — container running stale image without #2255 instrumentation>")
emit "phase_log" "{\"workspace_id\":\"$id\",\"container\":\"$container\",\"raw\":$(python3 -c "import json,sys; print(json.dumps(sys.argv[1]))" "$phase_log")}"
fi
done
fi
emit "run_completed" "{\"elapsed_secs\":$ELAPSED_SECS,\"pm_id\":\"$PM_ID\",\"child_id\":\"$CHILD_ID\"}"
cat <<EOF >&2
=========================================
Measurement complete. (RFC #2251 / Issue 4 repro)
Mode: $MODE
Coordinator template: $PM_TEMPLATE
Child template: $CHILD_TEMPLATE
Model: ${MODEL:-<template default>}
Coordinator response: ${ELAPSED_SECS}s
PM workspace: $PM_ID
Child workspace: $CHILD_ID
=========================================
Interpretation:
ELAPSED < 60 → Synthesis fast; not informative about platform bounds.
Re-run with SYNTHESIS_DEPTH=8 for longer synthesis.
60 <= ELAPSED < 300 → Within DELEGATION_TIMEOUT. Doesn't prove or refute
Issue 4 — HTTP-level timeout would be sufficient.
ELAPSED >= 300 → BUG CONFIRMED IF heartbeat_trace shows no platform-side
transition. Coordinator ran past DELEGATION_TIMEOUT without
any platform ceiling kicking in — exactly the gap V1.0
plans to close with MAX_TASK_EXECUTION_SECS.
curl_failed_or_timed_out → \$A2A_TIMEOUT exceeded. Coordinator likely hung
or synthesis is just very slow.
EOF

View File

@ -0,0 +1,107 @@
# Platform tool registry
Single source of truth for every tool the platform exposes to agents
(A2A delegation, hierarchical memory, broadcast, introspection).
## Why this exists
Pre-#2240, three places independently declared each tool:
1. **MCP server** (`workspace/a2a_mcp_server.py`) — the `TOOLS` JSON list
2. **LangChain `@tool` wrappers** (`workspace/builtin_tools/{delegation,memory}.py`)
3. **Agent-facing system-prompt docs** (`workspace/executor_helpers.py`)
Adding a tool to one and forgetting the others happened repeatedly. The
canonical case: `send_message_to_user` was registered in MCP TOOLS but
the executor_helpers doc string never mentioned it, so agents saw the
tool as available but had no usage guidance — a silent capability
regression.
## What the registry does
`registry.py` defines each tool ONCE as a frozen `ToolSpec`:
```python
ToolSpec(
name="delegate_task",
short="Delegate a task to a peer workspace via A2A and WAIT for the response.",
when_to_use="Use for QUICK questions and small sub-tasks where you can afford to wait inline...",
input_schema={...}, # JSON Schema, consumed by MCP server
impl=tool_delegate_task, # the actual coroutine
section="a2a", # which prompt section it belongs to
)
```
Adapters consume specs; no hardcoded names anywhere else:
- **MCP server** builds its `TOOLS` list from `_PLATFORM_TOOL_SPECS` at import time
- **LangChain `@tool` wrappers** read `name=spec.name` from the registry
- **Doc generator** (`executor_helpers._render_section()`) produces the
system-prompt block from `spec.short` (bullet) + `spec.when_to_use`
(heading + paragraph)
## CLI subprocess block — special case
Non-MCP runtimes (ollama, custom subprocess adapters) use a separate
hand-maintained block in `executor_helpers._A2A_INSTRUCTIONS_CLI` because
the CLI subcommand vocabulary (`peers`, `delegate`, `status`, `info`)
differs from the MCP tool names (`list_peers`, `delegate_task`, etc.).
Auto-generation would lose the readable invocation syntax.
Alignment is enforced via `_CLI_A2A_COMMAND_KEYWORDS` (in
`executor_helpers.py`): every a2a-section spec must be keyed there with
either a CLI subcommand keyword OR an explicit `None` if the tool is
intentionally not exposed via subprocess (e.g.
`send_message_to_user` because its structured `attachments` field
doesn't survive positional-arg shell invocation).
## Tests that catch drift
`workspace/tests/test_platform_tools.py`:
| Test | What it catches |
|---|---|
| `test_mcp_server_registers_every_registry_tool` | MCP TOOLS list out of sync with registry |
| `test_mcp_tool_descriptions_match_registry_short` | hand-edited MCP description that drifted |
| `test_mcp_tool_input_schemas_match_registry` | schema duplicated in server file |
| `test_a2a_instructions_text_includes_every_a2a_tool` | doc generator missed a tool |
| `test_old_pre_rename_names_not_present_in_docs` | stale name leaked back in |
| `test_a2a_mcp_instructions_match_snapshot` | rendered shape (bullet ordering, headings, footers) drifted |
| `test_a2a_cli_instructions_match_snapshot` | CLI block edited in a way that changes shape |
| `test_hma_instructions_match_snapshot` | HMA section drifted |
| `test_cli_keyword_mapping_covers_every_a2a_tool` | tool added to registry without a CLI mapping decision |
| `test_cli_keyword_substrings_appear_in_cli_block` | CLI keyword in the mapping but missing from the doc block |
The snapshot files at `workspace/tests/snapshots/*.txt` are LF-pinned
in `.gitattributes` so a Windows contributor with `core.autocrlf=true`
doesn't get mysterious test failures.
## Adding a new tool
1. Append a `ToolSpec(...)` to `TOOLS` in `registry.py`.
2. Add the LangChain `@tool` wrapper in `workspace/builtin_tools/`
(the wrapper body just calls `spec.impl`).
3. Update `_CLI_A2A_COMMAND_KEYWORDS` in `executor_helpers.py` — set the
value to the CLI subcommand keyword, or to `None` if the tool isn't
exposed via the subprocess interface.
4. Regenerate snapshots — see the comment block at the top of
`workspace/tests/test_platform_tools.py` for the one-liner.
5. Run `pytest workspace/tests/test_platform_tools.py --no-cov`.
## Renaming a tool
Edit `name` in `registry.py` only. Then:
1. The MCP TOOLS list rebuilds automatically.
2. The doc generator regenerates automatically (snapshots will fail
the diff — regenerate them).
3. Search `workspace/` for the old literal in case a non-adapter
consumer (tests, plugin code) hardcoded the old name; update those.
4. Update any `_CLI_A2A_COMMAND_KEYWORDS` key + the literal substring
in `_A2A_INSTRUCTIONS_CLI` if applicable.
## Removing a tool
Delete the `ToolSpec` and the `_CLI_A2A_COMMAND_KEYWORDS` key. Adapters
and doc generators stop registering it automatically; the structural
tests prevent stale references from surviving.